Re: st: sample adjustment by substitution instead of weighting

[Dirk Enzmann <dirk.enzmann@uni-hamburg.de>]

Thank you very much Steve for your elaborate answer - it is very 
helpful, indeed!

Dirk

On behalf of Steve I include his answer in reply to
http://www.hsph.harvard.edu/cgi-bin/lwgate/STATALIST/archives/statalist.0904/date/article-1134.html
here because at the moment he can't send mails to the list:

---------------------------------------------------------------------
Dirk-
I've never heard of this procedure. There is some basis for thinning a 
sample randomly to meet sampling goals,  and substitutions for missing 
observations are also practiced, but you are not describing either of these.

The process of exclusion and duplication will destroy the ability of the 
sample to estimate anything but the characteristics that are being 
matched--but those are already known!  For instance, the sample cannot 
estimate without bias the means of  other variates. For the matched 
characteristics, the sample will not permit estimation of SD's or 
quantiles.   Moreover, no standard errors or confidence intervals can be 
computed for anything, because the exclusions and duplication have 
artificially reduced the variability in the sample.

To better match the sample estimates to known population 
characteristics, I know of only three procedures: 1) post-stratification 
; 2) sample raking, which is an extension; and 3) generalized regression 
estimation (GREG).

The exclusions and duplication are naive attempts to re-weight the 
sample.  However they completely destroy it. So, no this is not actual 
practice. The only discussion of something similar I've read is in Lohr 
(1999, Sampling: Design and Analysis, Duxbury, p 463)  gives the 
reference to  Neyman J. 1934. On the two different methods of the 
representative method: The method of stratified sampling and the method 
of purposive selection. J. Royal Statistical Society 197: 558-606. Here 
is the quote from her book:

"Neyman's paper pretty much finished off the idea that results from 
purposive samples could be generalized to the population.  He presented 
an example of the purposive sample taken by Gini and Galvani in the late 
1920's. Gini and Galvani chose 29 districts that gave the averages of 
all 214 districts in the 1921 Italian census, on a dozen variables. But 
Neyman showed that all statistics other than the average values of the 
controls showed a violent contrast between the sample and the whole 
population."

Of course, Gini and Galvani only excluded, but did not duplicate, they 
only excluded.  So the procedure has long been discredited.

-Steve
---------------------------------------------------------------------

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/